Hephaestus: Improving Fundamental Agent Capabilities of Large Language Models through Continual Pre-Training